AI Automation Engineer
AI Automation Engineer
Therabot makes a robot that massages your legs. Behind the product is a company deliberately built to run on AI systems — and we are looking for the person who builds and runs those systems.
If you have been quietly building agents at night because you can't help yourself, and you have been waiting for somebody to hand you a real business to point that at: this is that job.
Be intentional about applying. The application ends by asking what you have built on these platforms specifically — agents, skills, workflows, integrations you have actually run. Have that ready before you start.
The idea behind the role
Most work falls into three layers. Doing it. Reviewing it. Deciding on it.
Our view is that the first layer is increasingly machine work. The people here should be reviewing that work or making the call on it — and the better the systems get, the less of even that is needed.
Doing
Pulling the numbers, drafting the reply, reconciling the transactions, building the report. Repetitive execution — the layer you will be automating.
Reviewing
A short daily pass over what the machine did — is it running, is it right, is anything stuck waiting on a person. Minutes, not hours.
Deciding
The handful of calls that genuinely need human judgment. These should arrive pre-framed: here are the three things worth your attention, yes or no.
What the job actually is
You will take a business function that currently eats a person's day, and turn it into a system that runs itself and asks for a human only when it genuinely should. Then the next one. Then the one after that.
- Build the systems. Agents, pipelines, scheduled jobs, and integrations across our real stack — Shopify, Meta, Google, email, accounting, and our own internal tools.
- Make them know when they are unsure. This is the hard part and the part we care most about. An agent that fails loudly is useful. An agent that fails confidently and quietly is worse than no agent at all.
- Build the trend layer. A snapshot that looks fine every week can be three percent wrong every week. Drift and silence are the two failures a human spot-check will never catch on its own.
- Turn every correction into a permanent fix. When a human catches something, that class of error should never reach a human again. If we are catching the same thing in month three that we caught in week one, the system is not learning.
- Delete the check-ins you created. Your scorecard is not how many systems need watching. It is how few.
The measure of success: more of the business covered, less human attention required. Those two numbers should move in opposite directions month over month, and you should be the one who can prove they did.
Who this is for
You will love this if
- You are already building on OpenClaw or Hermes today
- You already build agents nobody asked you to build
- You have gone back and made something better after it already worked
- You get an actual kick out of deleting your own recurring tasks
- You want to own a hard problem end to end rather than manage people
- You are fine being handed an outcome and left to decide the how
- You want your work touching a real product with real customers and real money
You will hate this if
- You have never used OpenClaw or Hermes and do not want to
- You want a spec, a ticket queue, and a defined scope
- You want to be told which tools and which architecture to use
- Your best work is demos rather than things that survive six months
- You are looking for a title and a team to manage
- You need a large engineering org around you to do your best work
- You would find a small hardware company's messiness frustrating rather than fun
What we care about, and what we don't
We do not care about your degree, your job titles, whether you call yourself an engineer, or how long your résumé is. Several of the best people for this job are not engineers by title at all — they are operations people, ex-freelancers, ex-solo-founders, and the person at their last company who built the internal tools nobody asked for and everybody ended up depending on.
We do care about one thing above all: what you have built that somebody else relied on, and what happened when it broke. That is the whole filter. Build something forty people use daily and you can skip every credential.
The deal
Compensation: negotiable salary. To be upfront: we are not offering equity at this time — this is a cash-only package. We would rather hear your number early than dance around it, and there is a question about it in the application.
What you should know going in: we are small, we are pre-launch on our first production run, and the work is real. You will not be building toys. You will also not be inheriting a mature system to maintain — you are building it from close to zero, which is either exactly what you want or exactly what you don't.
How hiring works
- Step 1 — the application below. About 15 minutes. It starts with how you work and how you think, and ends by asking you to show what you have actually built — have a link or a description ready. No cover letter, no résumé required.
- Step 2 — a real problem. If it is a fit, we send you a broken agent scenario and ask how you would diagnose it. Paid if it takes real time.
- Step 3 — a conversation with Aaron. Direct, technical, no panel.
- Step 4 — a small paid project. Real work, real scope, so we both find out what working together is like before either of us commits.
We read every application. We reply either way.
Start with the application.
About 15 minutes. It ends by asking what you have built on OpenClaw or Hermes — have that ready. Write your own answers; we can tell, and generic AI-written responses are the fastest way to a no. Rough and specific beats polished and vague.
Apply now →Questions before you apply? Email support@therabot.com with "Careers" in the subject line.